Clarify or Answer: Reinforcement Learning for Agentic VQA with Context Under-specification
CoA is an agentic framework that decides whether to answer a visual question directly or ask for clarification first. It uses a controller to decide the action, a reinforcement learning policy (GRPO-CR) to generate targeted questions, and an answering model. This method improves accuracy by strategically asking one clarification question when context is missing, showing better performance on both training and new datasets.